Research map / Federated learning and privacy / Personalized and heterogeneous FL

Scalable Federated One-Step Multi-View Clustering with Tensorized Regularization

AAAI 2025 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Federated medical and multimodal learning · 6 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SFOMVC-TR, a federated one-step multi-view clustering method that projects anchor graphs directly to cluster assignments via Lp,q sparse regularization and aligns client assignments on the server with tensor Schatten-p regularization.

Problem

Federated multi-view clustering where multi-view data are held by different organizations; existing FMVC methods need post-processing (e.g., K-means) to get labels, are costly on high-dimensional data, and struggle to capture complementary and consistent information while preserving privacy.

Approach

Combine anchor graphs for scalability with one-step clustering on each client (a sparse Lp,q-regularized projection from anchor graph to cluster assignments), and have the server fuse client assignments as a tensor with Schatten-p regularization, optimized via an ADMM-style federated scheme that shares only assignment matrices and scalar losses, never raw data or anchor graphs.

Key results

• On ORL, SFOMVC-TR reaches 100% ACC, NMI and PUR, versus 99.75% ACC for the strongest federated baseline TensorFMVC.
• On the large Reuters dataset, SFOMVC-TR achieves 72.79% ACC, 64.51% NMI and 77.22% PUR, while several baselines run out of memory.
• Ablation shows removing the tensor regularizer drops 3-sources ACC from 84.02% to 60.95%, and the two-step variant performs much worse than the one-step version on most datasets (e.g., NGs ACC 30.40% vs 99.00%).

Measured effects reported in the paper

Surprising findings

• The two-step variant (SVD plus K-means post-processing) collapses to around 30% ACC on NGs, far below the one-step version.

Limitations

• Lp,q regularization is solved only for p=2 with 0<q<=1 due to lack of a closed-form solution for general p and q.
• The final aggregation is only performed after convergence, and privacy is argued informally with a semi-honest model rather than formally proven.
• Several centralized and federated baselines fail with out-of-memory errors on larger datasets, limiting comparison.

Experimental setup

Datasets: ORL Yale BBCSport 3-sources WebKB UCI-digit NGs Reuters

Benchmarks: ORL Yale BBCSport 3-sources WebKB UCI-digit NGs Reuters

Metrics: ACC NMI PUR

Baselines: FastMICE MvLRSSC RMSL GMC FPMVS-CAG VKMC FCVFL FedMVL FMVC-IMK TensorFMVC

Builds on: MVSC anchor graph clustering NMF-based federated multi-view clustering (FedMVL) Lp,q regularization (Hu et al. 2017) tensor Schatten-p norm

Closest papers by meaning

• Bridging Gaps: Federated Multi-View Clustering in Heterogeneous Hybrid Views (NeurIPS 2024 · similarity 0.91)
• Efficient Federated Incomplete Multi-View Clustering (ICML 2025 · similarity 0.90)
• Graph Consistency and Diversity Measurement for Federated Multi-View Clustering (AAAI 2025 · similarity 0.90)
• S2MVTC: a Simple yet Efficient Scalable Multi-View Tensor Clustering (CVPR 2024 · similarity 0.89)
• An Effective and Secure Federated Multi-View Clustering Method with Information-Theoretic Perspective (ICML 2025 · similarity 0.89)
• Federated Incomplete Multi-view Clustering with Globally Fused Graph Guidance (ICML 2025 · similarity 0.88)